Amazon researchers have developed BASE TTS, the largest text-to-speech model to date, with 980 million parameters and trained on 100,000 hours of speech. This model demonstrates emergent abilities, handling complex linguistic challenges such as compound nouns, emotional expressions, foreign words, and paralinguistics more effectively than previous models. The medium-sized version of BASE TTS, with 400 million parameters, notably exhibited a significant improvement in these areas. The model’s streamable nature allows for real-time, low-bitrate speech generation, with the potential to enhance accessibility features. Although impressive, the model remains experimental, and its source data has not been published due to potential misuse by bad actors.
Read more at TechCrunch…
